x

2026-06-10 10:52:32 +08:00
parent 0ba03bc3f8
commit 50c52d7d5f
3 changed files with 106 additions and 72 deletions
@@ -0,0 +1,20 @@
+# Docker 环境变量（按需修改）
+# 复制为 .env 后使用: cp env.example .env
+
+# ---- GPU 架构（按你的显卡修改）----
+# gfx1201 = RX 9070 / 9070 XT / 9070 GRE
+# gfx1200 = RX 9060 XT / 9060 XT LP
+# gfx1100 = RX 7900 XTX / XT / GRE
+# gfx1101 = RX 7800 XT / 7700 XT
+# gfx1030 = RX 6950 / 6900 / 6800 系列
+ARCH=gfx1201
+
+# ---- 目录挂载 ----
+INPUT_DIR=./data/input
+OUTPUT_DIR=./data/output
+MODEL_DIR=./data/models
+MIOPEN_CACHE=./data/miopen
+
+# ---- 模型下载源 ----
+# huggingface（默认，需科学上网）或 modelscope（国内可用）
+MINERU_MODEL_SOURCE=modelscope
@@ -15,7 +15,7 @@ ARG TORCH_INDEX=https://download.pytorch.org/whl/rocm7.2
 ARG GIT_PROXY=http://127.0.0.1:8118

 # -- 国内镜像 ----------------------------------------------------------------
-ARG PIP_INDEX=https://pypi.tuna.tsinghua.edu.cn/simple
+ARG PIP_INDEX=https://mirrors.aliyun.com/pypi/simple

 # -- 环境变量 ---------------------------------------------------------------
 ENV DEBIAN_FRONTEND=noninteractive \
@@ -95,10 +95,10 @@ RUN python${PYTHON_VER} -m venv ${VENV} && \
    mkdir -p /root/.pip && \
    echo "[global]" > /root/.pip/pip.conf && \
    echo "index-url = ${PIP_INDEX}" >> /root/.pip/pip.conf && \
-    ${VENV}/bin/pip install --no-cache-dir -U pip setuptools wheel && \
+    ${VENV}/bin/pip install -U pip setuptools wheel && \
    # PyTorch ROCm wheels 在海外 CDN，通过代理下载
    export http_proxy=${GIT_PROXY} https_proxy=${GIT_PROXY} && \
-    ${VENV}/bin/pip install --no-cache-dir --pre \
+    ${VENV}/bin/pip install --pre \
        torch==2.11.0+rocm7.2 \
        torchvision \
        pytorch-triton-rocm \
@@ -120,7 +120,7 @@ RUN apt-get update && apt-get install -y --no-install-recommends \
 # ===========================================================================
 RUN if [ -d /opt/rocm/share/amd_smi ]; then \
        cp -r /opt/rocm/share/amd_smi /opt/amd_smi && \
-        cd /opt/amd_smi && ${VENV}/bin/pip install --no-cache-dir . --no-build-isolation && \
+        cd /opt/amd_smi && ${VENV}/bin/pip install . --no-build-isolation && \
        echo "amdsmi installed."; \
    else \
        echo "amdsmi source not found, skipping."; \
@@ -132,8 +132,8 @@ RUN if [ -d /opt/rocm/share/amd_smi ]; then \
 COPY aiter/ /opt/aiter/
 COPY flash-attention/ /opt/flash-attention/
 RUN set -ex && \
-    ${VENV}/bin/pip install --no-cache-dir -e /opt/aiter && \
-    ${VENV}/bin/pip install --no-cache-dir --no-build-isolation -e /opt/flash-attention && \
+    ${VENV}/bin/pip install -e /opt/aiter && \
+    ${VENV}/bin/pip install --no-build-isolation -e /opt/flash-attention && \
    ${VENV}/bin/python -c "import torch; v=torch.__version__; assert 'rocm' in v, f'PyTorch overwritten: {v}'; print('PyTorch OK:', v)"

 # ===========================================================================
@@ -141,7 +141,7 @@ RUN set -ex && \
 # ===========================================================================
 COPY vllm/ /opt/vllm/
 RUN set -ex && \
-    ${VENV}/bin/pip install --no-cache-dir -U \
+    ${VENV}/bin/pip install -U \
        "setuptools>=77.0.3" setuptools_scm setuptools_rust wheel && \
    cd /opt/vllm && \
    if [ -f csrc/mamba/mamba_ssm/selective_scan.h ]; then \
@@ -277,12 +277,12 @@ COPY scripts/cache_warmer.py /opt/cache_warmer.py
 # 阶段 8b：安装 vllm + 平台补丁 + 验证
 # ===========================================================================
 RUN set -ex && \
-    cd /opt/vllm && ${VENV}/bin/pip install --no-cache-dir -e . --no-build-isolation && \
-    ${VENV}/bin/pip install --no-cache-dir regex && \
+    cd /opt/vllm && ${VENV}/bin/pip install -e . --no-build-isolation && \
+    ${VENV}/bin/pip install regex && \
    ${VENV}/bin/python /opt/patch_vllm_platform.py && \
    ${VENV}/bin/python -c "import torch; v=torch.__version__; assert 'rocm' in v, f'PyTorch overwritten: {v}'; print('PyTorch OK:', v)" && \
    ${VENV}/bin/pip uninstall -y triton triton-rocm 2>/dev/null; \
-    ${VENV}/bin/pip install --no-cache-dir --force-reinstall \
+    ${VENV}/bin/pip install --force-reinstall \
        torch==2.11.0+rocm7.2 torchvision pytorch-triton-rocm \
        --index-url ${TORCH_INDEX} && \
    ${VENV}/bin/python -c "import vllm; print('vllm import OK:', vllm.__version__)" && \
@@ -292,13 +292,13 @@ RUN set -ex && \
 # 阶段 9：安装 MinerU + RDNA 适配补丁 + PyTorch 恢复
 # ===========================================================================
 RUN set -ex && \
-    ${VENV}/bin/pip install --no-cache-dir 'mineru[core]' && \
+    ${VENV}/bin/pip install 'mineru[core]' && \
    ${VENV}/bin/python /opt/apply_mineru_patches.py && \
    # mineru[core] 会把 ROCm PyTorch 替换成 CUDA 版 + 安装 CUDA triton
    # 必须先卸载 CUDA triton，再装回 ROCm 版
    ${VENV}/bin/pip uninstall -y triton triton-rocm 2>/dev/null; \
    export http_proxy=${GIT_PROXY} https_proxy=${GIT_PROXY} && \
-    ${VENV}/bin/pip install --no-cache-dir --force-reinstall \
+    ${VENV}/bin/pip install --force-reinstall \
        torch==2.11.0+rocm7.2 torchvision pytorch-triton-rocm \
        --index-url ${TORCH_INDEX} && \
    ${VENV}/bin/python -c "import triton.language; print('triton OK')" && \
@@ -4,41 +4,7 @@
 # =============================================================================

 services:
-  # --- CLI 模式（默认）---
-  mineru:
-    image: mineru-rocm:7.2.1
-    build:
-      context: .
-      dockerfile: Dockerfile
-      network: host
-      args:
-        ARCH: gfx1201
-        GIT_PROXY: http://127.0.0.1:8118
-    profiles: ["cli"]
-    container_name: mineru-rocm
-    stdin_open: true
-    tty: true
-    ipc: host
-    devices:
-      - /dev/kfd
-      - /dev/dri
-    security_opt:
-      - seccomp=unconfined
-    group_add:
-      - video
-    environment:
-      - MINERU_MODEL_SOURCE=${MINERU_MODEL_SOURCE:-huggingface}
-      - HF_HUB_CACHE=${HF_HUB_CACHE:-/opt/models/huggingface}
-      - MODELSCOPE_CACHE=${MODELSCOPE_CACHE:-/opt/models/modelscope}
-    volumes:
-      - ${INPUT_DIR:-./data/input}:/data/input:ro
-      - ${OUTPUT_DIR:-./data/output}:/data/output
-      - ${MODEL_DIR:-./data/models}:/opt/models
-      - ${MIOPEN_CACHE:-./data/miopen}:/root/.cache/miopen
-    command: bash
-    restart: "no"
-
-  # --- WebUI（Gradio）---
+  # --- WebUI 前端 ---
  gradio:
    image: mineru-rocm:7.2.1
    profiles: ["gradio"]
@@ -48,13 +14,6 @@ services:
    ipc: host
    ports:
      - "10002:7860"
-    devices:
-      - /dev/kfd
-      - /dev/dri
-    security_opt:
-      - seccomp=unconfined
-    group_add:
-      - video
    environment:
      - MINERU_MODEL_SOURCE=${MINERU_MODEL_SOURCE:-huggingface}
      - HF_HUB_CACHE=${HF_HUB_CACHE:-/opt/models/huggingface}
@@ -63,33 +22,88 @@ services:
      - ${INPUT_DIR:-./data/input}:/data/input:ro
      - ${OUTPUT_DIR:-./data/output}:/data/output
      - ${MODEL_DIR:-./data/models}:/opt/models
-      - ${MIOPEN_CACHE:-./data/miopen}:/root/.cache/miopen
-    command: ["mineru-gradio --server-name 0.0.0.0 --server-port 7860"]
+    command:
+      [
+        "mineru-gradio",
+        "--server-name", "0.0.0.0",
+        "--server-port", "7860",
+        "--api-url", "http://mineru-router:8000",
+      ]
+    depends_on:
+      - router

-  # --- API 服务 ---
-  api:
+  # --- 双卡 Worker（GPU 算力，无 profile，始终可用）---
+  worker0:
    image: mineru-rocm:7.2.1
-    profiles: ["api"]
-    container_name: mineru-api
+    container_name: mineru-worker0
+    stdin_open: true
+    tty: true
+    ipc: host
+    devices:
+      - /dev/kfd
+      - /dev/dri
+    security_opt:
+      - seccomp=unconfined
+    group_add:
+      - video
+    environment:
+      - HIP_VISIBLE_DEVICES=0
+      - MINERU_MODEL_SOURCE=${MINERU_MODEL_SOURCE:-huggingface}
+      - HF_HUB_CACHE=${HF_HUB_CACHE:-/opt/models/huggingface}
+      - MODELSCOPE_CACHE=${MODELSCOPE_CACHE:-/opt/models/modelscope}
+    volumes:
+      - ${INPUT_DIR:-./data/input}:/data/input:ro
+      - ${OUTPUT_DIR:-./data/output}:/data/output
+      - ${MODEL_DIR:-./data/models}:/opt/models
+      - ${MIOPEN_CACHE:-./data/miopen}:/root/.cache/miopen
+    command: ["mineru-api", "--host", "0.0.0.0", "--port", "8001"]
+
+  worker1:
+    image: mineru-rocm:7.2.1
+    container_name: mineru-worker1
+    stdin_open: true
+    tty: true
+    ipc: host
+    devices:
+      - /dev/kfd
+      - /dev/dri
+    security_opt:
+      - seccomp=unconfined
+    group_add:
+      - video
+    environment:
+      - HIP_VISIBLE_DEVICES=1
+      - MINERU_MODEL_SOURCE=${MINERU_MODEL_SOURCE:-huggingface}
+      - HF_HUB_CACHE=${HF_HUB_CACHE:-/opt/models/huggingface}
+      - MODELSCOPE_CACHE=${MODELSCOPE_CACHE:-/opt/models/modelscope}
+    volumes:
+      - ${INPUT_DIR:-./data/input}:/data/input:ro
+      - ${OUTPUT_DIR:-./data/output}:/data/output
+      - ${MODEL_DIR:-./data/models}:/opt/models
+      - ${MIOPEN_CACHE:-./data/miopen}:/root/.cache/miopen
+    command: ["mineru-api", "--host", "0.0.0.0", "--port", "8002"]
+
+  router:
+    image: mineru-rocm:7.2.1
+    container_name: mineru-router
    stdin_open: true
    tty: true
    ipc: host
    ports:
      - "8000:8000"
-    devices:
-      - /dev/kfd
-      - /dev/dri
-    security_opt:
-      - seccomp=unconfined
-    group_add:
-      - video
    environment:
      - MINERU_MODEL_SOURCE=${MINERU_MODEL_SOURCE:-huggingface}
-      - HF_HUB_CACHE=${HF_HUB_CACHE:-/opt/models/huggingface}
-      - MODELSCOPE_CACHE=${MODELSCOPE_CACHE:-/opt/models/modelscope}
    volumes:
      - ${INPUT_DIR:-./data/input}:/data/input:ro
      - ${OUTPUT_DIR:-./data/output}:/data/output
-      - ${MODEL_DIR:-./data/models}:/opt/models
-      - ${MIOPEN_CACHE:-./data/miopen}:/root/.cache/miopen
-    command: ["mineru-api --host 0.0.0.0 --port 8000"]
+    command:
+      [
+        "mineru-router",
+        "--api-urls",
+        "http://mineru-worker0:8001,http://mineru-worker1:8002",
+        "--host", "0.0.0.0",
+        "--port", "8000",
+      ]
+    depends_on:
+      - worker0
+      - worker1